app.py

# import part
import streamlit as st
from transformers import pipeline
from PIL import Image

# Set global caching options for Transformers
from transformers import set_caching_enabled
set_caching_enabled(True)

# function part with caching for better performance
@st.cache_resource
def load_image_captioning_model():
    return pipeline("image-to-text", model="sooh-j/blip-image-captioning-base")

@st.cache_resource  
def load_text_generator():
    return pipeline("text-generation", model="TinyLlama/TinyLlama-1.1B-Chat-v1.0")

@st.cache_resource
def load_tts_model():
    return pipeline("text-to-speech", model="HelpingAI/HelpingAI-TTS-v1")

# img2text - Using the original model with more constraints
def img2text(image):
    # Load the model (cached)
    image_to_text = load_image_captioning_model()
    
    # Strongly limit output length for speed
    text = image_to_text(image, max_new_tokens=15)[0]["generated_text"]
    return text

# text2story - Much more constrained for speed
def text2story(text):
    # Load the model (cached)
    generator = load_text_generator()
    
    # Very brief prompt to minimize work
    prompt = f"Short story about {text}: Once upon a time, "
    
    # Very constrained parameters for maximum speed
    story_result = generator(
        prompt,
        max_new_tokens=60,  # Much shorter output
        num_return_sequences=1,
        temperature=0.7,
        top_k=10,  # Lower value = faster
        top_p=0.9,  # Lower value = faster
        do_sample=True
    )
   
    # Extract and clean text
    story_text = story_result[0]['generated_text']
    story_text = story_text.replace(prompt, "Once upon a time, ")
    
    # Find a natural ending point
    last_period = story_text.rfind('.')
    if last_period > 30:  # Ensure we have at least some content
        story_text = story_text[:last_period + 1]
    
    return story_text

# text2audio - Minimal text for faster processing
def text2audio(story_text):
    try:
        # Load the model (cached)
        synthesizer = load_tts_model()
        
        # Aggressively limit text length to speed up TTS
        max_chars = 200  # Much shorter than before
        if len(story_text) > max_chars:
            last_period = story_text[:max_chars].rfind('.')
            if last_period > 0:
                story_text = story_text[:last_period + 1]
            else:
                story_text = story_text[:max_chars]
        
        # Generate speech
        speech = synthesizer(story_text)
        return speech
        
    except Exception as e:
        st.error(f"Error generating audio: {str(e)}")
        return None

# Streamlined main UI
st.set_page_config(page_title="Image to Story", page_icon="📚")
st.header("Image to Audio Story")

# Add info about processing time
st.info("Note: Processing may take some time as the models are loading. Please be patient.")

# Cache the file uploader state
if "uploaded_file" not in st.session_state:
    st.session_state["uploaded_file"] = None
    
uploaded_file = st.file_uploader("Select an Image...", key="file_uploader")

# Process the image if uploaded
if uploaded_file is not None:
    st.session_state["uploaded_file"] = uploaded_file
    
    # Display the uploaded image
    st.image(uploaded_file, caption="Uploaded Image", use_column_width=True)
    
    # Convert to PIL image
    image = Image.open(uploaded_file)
    
    # Optional processing toggle to let user decide
    if st.button("Generate Story and Audio"):
        col1, col2 = st.columns(2)
        
        # Stage 1: Image to Text with minimal output
        with col1:
            with st.spinner('Captioning image...'):
                caption = img2text(image)
            st.write(f"**Caption:** {caption}")
        
        # Stage 2: Text to Story with minimal length
        with col2:
            with st.spinner('Creating story...'):
                story = text2story(caption)
            st.write(f"**Story:** {story}")
        
        # Stage 3: Audio with minimal text
        with st.spinner('Generating audio...'):
            speech_output = text2audio(story)
        
        # Display audio immediately
        if speech_output is not None:
            try:
                if 'audio' in speech_output and 'sampling_rate' in speech_output:
                    st.audio(speech_output['audio'], sample_rate=speech_output['sampling_rate'])
                elif 'audio_array' in speech_output and 'sampling_rate' in speech_output:
                    st.audio(speech_output['audio_array'], sample_rate=speech_output['sampling_rate'])
                elif 'waveform' in speech_output and 'sample_rate' in speech_output:
                    st.audio(speech_output['waveform'], sample_rate=speech_output['sample_rate'])
                else:
                    # Try any array-like data
                    for key, value in speech_output.items():
                        if hasattr(value, '__len__') and len(value) > 1000:
                            sample_rate = speech_output.get('sampling_rate', speech_output.get('sample_rate', 24000))
                            st.audio(value, sample_rate=sample_rate)
                            break
                    else:
                        st.error("Could not find audio data in the output")
            except Exception as e:
                st.error(f"Error playing audio: {str(e)}")
        else:
            st.error("Audio generation failed")